Agent observability is the practice of using telemetry to understand an agent's activity, including its interactions, execution paths, and observed outcomes.
This operational definition concerns what can be learned from the instrumented workflow. It does not imply complete access to every component or to a model's internal reasoning. An explanation generated by the agent and a record emitted by an executor have different evidentiary roles.
What the observations can describe
Relevant observations may identify an agent invocation, a selected tool, a policy decision, an attempted operation, or a returned result. Their availability depends on the runtime and components that emit them. A useful description names which part of the workflow produced each record and which question it can answer.
OpenTelemetry's semantic conventions for GenAI agent spans distinguish a client invocation of a remote agent from invocation within the same process. These conventions are marked Development. They illustrate different instrumentation boundaries; they do not establish that a deployment emits either span or observes the remote agent's internal work.
Preserve the meaning of each source
An event reporting a policy decision describes the evaluator's result. An executor event can describe dispatch or local handling. A destination record may describe acceptance or a resource change. Combining those records should preserve their distinctions rather than turn every successful response into a completed task.
A session identifier can group related work without identifying the attempt that caused an effect. Likewise, an absent event can reflect an uninstrumented path or incomplete collection. State the gap before drawing a conclusion about activity that was not observed.
Choose the evidence needed for the question. Capturing full prompts, outputs, or tool arguments is not automatically necessary; controlled references and selected attributes may suffice. More collected content does not by itself make the interpretation more reliable.
How observability differs from governance
AI agent governance defines authority, policy, controls, and responsibilities. Observations can supply context to those controls while work is running, as well as support later review. The existence of telemetry does not alone establish that a decision constrained execution.
The analysis of observing versus governing an agent examines that additional control claim. AI observability has a broader scope that can include model and application behavior without an agent workflow.
The AI agent audit trail guide explains how to assemble observations into reviewable evidence with explicit identities, stages, sources, and limitations.